Autonomous driving in a city? We're '95% of the way there'
Goal is 'Level 5', where vehicle reacts like human-driven one
Universities, carmakers, local authorities and tech and insurance firms are involved in an array of part-government-sponsored self-driving vehicle pilots across the UK.
More than £20m is being funnelled into projects to build and refine autonomous technology, understand human-vehicle interaction, and assess risk.
It’s an era described as the "Kitty Hawk" moment for robot cars – a reference to the moment in 1903 when the Wright Brothers made their first powered flight in North Carolina. What followed was a blossoming of ideas in aircraft design, technology and new companies.
Home shopping via robot car
In June, British supermarket retailer Ocado jumped into the fray, partnering with Oxford University spin-out Oxbotica on a 10‑day pilot of CargoPod.
CargoPod is an electric, self-driving vehicle that ferried goods for 100 customers on a 3km route around Greenwich at a top speed of 5mph (8kmph).
CargoPod was based on technology in one of those government-supported projects – GATEway, whose bug-like-shuttles have been trundling around Greenwich since October 2015 running on Oxbotica’s Selenium autonomous operating system and its Caesium Microsoft-Azure-based shuttle management system.
Ocado’s debut marked a turning point in Britain’s many driverless trials: an actual, potential end user as opposed to those making the technology.
As the UK’s only purely online supermarket, Ocado has prioritised technology investment in a way not common among others. Ocado’s entire web shop, order management and supply chain runs on cloud – AWS for compute and Google for data – which forms its so‑called Ocado Smart Platform. Developing and expanding this is Ocado Technology, a unit of 1,000 people spanning the UK, Spain, Poland and Bulgaria.
CargoPod: Ocado’s toe-in-the water on driverless vehicles. Pic: Gavin Clarke
The division operates on a philosophy of "10X" – seeking projects that improve Ocado revenue by a factor of 10. Ocado claims more than 200 patent applications across 50 separate inventions in more than 15 jurisdictions. One of the latest is a 4G‑based communication network using unlicensed 5GHz for point-to-point communications of up to 10 times a second at its latest warehouse in Andover.
Ocado is keen to not only use but also license the technology for use in vehicle-to-vehicle communications.
David Sharp, Ocado's head of 10x, told The Reg: “Ocado likes to surf the edge of practical technology and in the 10X department. We want to find all the things that are on the way of becoming that. This [driverless project] is one of them. It is a given we will be using this sort of technology – the timing is 'when it becomes practical'.”
And therein lies the challenge. What is “practical” for an enterprise like Ocado?
The national and gadget press were predictably left awed at the prospect of driverless Ocado deliveries. But, first things first.
There has been a lot of hype about driverless techin the UK and US as carmaker after carmaker has teamed up with tech firms and universities on technology and pilots.
The goal is level-five automation – where the vehicle performs as if being driven by a human. These categories were laid out by the US Department of Transportation’s National Highway Traffic Safety Administration (NHTSA, here) in 2013 but updated three years later to accept the guidelines of the SAE International (initially established as the Society of Automotive Engineers), a group spanning manufacturers in the aerospace, automotive and commercial-vehicle industries (PDF here).
Oxbotica chief executive Graeme Smith told The Reg that the goal is for level four or five autonomous driving in a city environment by 2019 – another two years away.
As good as human driving
Currently, says Smith, the firm is "95 per cent there"; that final five per cent is “complicated”.
CargoPod’s shuttle relatives on the GATEway project have been designed to operated at "level four" – the system can perform all the safety-critical driving functions and monitor road conditions for the whole trip. In reality, however, they are being used in level-two mode – with a driver who doesn’t intervene unless something goes wrong.
Critically, for level four, the vehicle must run within its “operational design domain”. Translated: it must be prepared to handle every driving scenario it knows or has been designed for – meaning, by the book, level four does not cover either unexpected situations or unknown terrain.
Smith said: “Getting to 95 per cent is relatively easy – getting to the final five per cent is much more complicated and a lot of that comes down to being here [in Greenwich] – collecting data, taking it away and figuring out what happened.
“Somebody running out in front of a vehicle is predicable, but when you choose to overtake on an A or a B road, that has a lot more risk. Those are the types of things we look at in more detail, about how you’d take those decisions.”
Get this disk to the Microsoft DC - it's too much for the cloud
CargoPod generated a stunning 4TB a day during the course of its 10‑day pilot. This was uploaded to Azure for analysis by disk to Microsoft’s data centre in Dublin because it’s too much for the network. That was a simple and slow campus-like environment dealing with people and vehicles.
By comparison, Ocado delivery vans today generate an average of 5MB of data based on hundred of events per day from sensors on board each. Data includes wheel speed, length of time a van was parked, GPS, and customer postcodes to – Ocado says – learn customer patterns. That data is fed into the OSP cloud that calculates and calibrates the ideal routes by crunching four million journey changes a second. The Ocado maxim is "no delivery should make a loss".
Geni: the vehicle driving Oxbotica’s autonomous ambition (Photo by Oxbotica)
The kinds of scenarios alluded to by Smith in his talk of overtaking will start to be played out in September, when a fleet of six Oxbotica vehicles will be released on the M40 motorway linking Oxford and London under the £13m DRIVEN project. These vehicles will operate at level four with humans onboard to intervene.
Oxbotica has not released details of the vehicles, but they will likely be modelled on its Smart-like Geni two-seater, which is the basis GATEway shuttle and another project – LUTZ Pathfinder testing autonomous vehicles in pedestrianised areas, conducted in Milton Keynes.
The Oxford-to-London run takes place in fast-moving traffic where high-risk scenarios like overtaking will play out – an act where decisions and actions can produce profound consequences. There will also be dynamic situations such as roadworks, where the vehicle is forced to leave a pre-marked course track and drive down temporary lanes.
“We’re getting there in small steps – it’s not a linear trajectory, it’s a spiral trajectory and getting better at certain things in certain environments,” Smith said of the experience.
Machine leaner driver
Professor Nick Reed, academy director of transport research body TRL, which is participating in GATEway, likened the process to a newly qualified human learner driver.
“You can compare it to a human driver, when you are learning to drive on a quiet suburban road until you are ready to step up to busy environments like roundabouts,” he said.
“Not even a human driver will have seen everything and there will be a situation that surprises you. The automated systems are going through that process and then they try to share that experience across every vehicle using the system.
“They will get to a situation where they need to let these vehicles operate in the wild, but they will be very cautious in that they slow down for more experiences than you’d expect them to. As you develop and improve their ability to detect and respond, you will be able to improve their behavior,” Reed told The Reg.
Learning is one thing, but cost of manufacturing of these vehicles and making them affordable is another practical issue for firms like Ocado. Smith reckons the biggest hurdle to self-driving vehicles is not intelligence but cost of sensors needed.
Take light detection and ranging (LiDAR) – a technology synonymous with self-driving vehicles. LiDAR sensors emit pulsed laser beams so the device knows where it is and its distance from objects, giving the vehicle 360-degree proximity vision. Put LiDAR on a self-driving vehicle and it should be able to sense what’s nearby and the range – down to millimeters.
Google established the popular image and adoption of LiDAR during its self-driving tests as a large cylindrical device mounted atop a vehicle. These rotating devices were made by Velodyne and priced at $64,000. There has been a move to cheaper, more streamlined sensors including non-rotating, solid-state LiDAR with Waymo – the Alphabet spinout carrying the baton of those early Google self-driving tests. These will be 90 per cent cheaper, Waymo says, but will still come in at $6,700. Velodine has promised a solid state LiDAR that, at volume, could be as low as $50 per sensor. Oxbotica’s Greenwich vehicles use rotating LiDAR.
Joe Kempton, autonomous vehicle analyst for Canalys, said: “Across the board, prices of all sensors in vehicles are coming down and that’s been driven by demand and supply, as more automotive OEMs are demanding their highly advanced sensors.”
The ball is therefore in the court of the carmakers – the likes of Audi and Volvo – to ramp up production of vehicles, thus leading to more supply and driving down prices. Oxbotica’s Smith reckons this will happen around 2020-2021.
“The automotive market is going to lead and break those barriers first,” he told The Reg. “Once the auto companies begin to push products we will be ready. The kinds of vehicles we are looking at today will be ready after the car companies have launched.”
GATEway shuttle is Oxbotica’s slow-moving technology and socialisation experiment (Photo by Gavin Clarke)
But there's another practical question for any potential like Ocado.
What self-driving technology should you bet on?
If autonomous vehicles are just computers on wheels, and if they follow the history of the computer industry, expect consolidation and shakeout.
Detecting a winner in this Kitty Hawk throng will be difficult. While first to powered flight, the Wright brothers failed to make a commercial success of their pioneering work. Timber company executive William Boeing, meanwhile, was so captivated by the sight of a flying machine in 1909 that he founded the Boeing Company in 1916, which proceeded to dominate civilian and military aircraft manufacturing.
Back on the ground and in the present day, the number of companies authorized to test autonomous vehicles in California, home of Silicon Valley, is today numbered at 36 – up from 11 last year.
That list is comprised of carmakers, component manufacturers and tech firms – from chips to online search and smart phones.
In the UK, the British government last year gave its blessing to eight projects working on a number of aspects – vehicle systems, testing tools, user acceptance and hardware – to retrieve and analyse vehicle data supposedly in real-time.
The Royal Borough of Greenwich is home to two autonomous vehicle pilots: GATEway, an £8m project dating from October 2015, and MOVE‑UK, a three-year project that’s part of the UK government’s £20m series of announcements last year.
MOVE-UK hasn’t revealed much of its technology, but the project is looking at data: devices mounted on five council-owned Land Rover Discoveries are capturing data on how the driver responded and how the autonomous vehicle would have acted.
In such a competitive environment, Oxbotica is coy about detailing its technology, although it claims 85 pieces of Intellectual Property in its Selenium self-driving system, which it licenses whole or in part to tier-one automakers.
CargoPod, which uses Selenium, employs a visual rather than GPS navigation: a series of images are compared to incoming video streams from cameras on the van for orientation. Oxbotica eschewed GPS, saying it means they aren’t dependent on third-party map providers.
Image processing is performed on what Oxbotica called a “high-end, off-the-shelf” laptop running an Intel Xeon 12‑core processor and Ubuntu Linux. Selenium is written in C/C++ and is platform agnostic, Oxbotica says.
Stepping out from the UK you have Nvidia, which is pushing a GPU stack and rallying vehicle makers against PC chip giant, Intel. Nvidia’s Drive PX uses its Tegra Systems on a Chip with up to eight Cortex cores and 512 CUDA cores. The software is Nvidia’s written using C/C++ and using GPU-accelerated Deep Learning SDKs for Caffe, CNTK, TensorFlow, Theano, and Torch. Nvidia has an impressive number of carmakers including Toyota, Volvo, Tesla, Audi, Mercedes-Benz and car systems maker Bosch – that’s part of DRIVE‑UK with Jaguar Land Rover.
Which autonomous vehicle platform would you choose? (Photo by Oxbotica)
Intel has been working with BMW and Delphi and spent $15bn this year on Mobileye, developing vision-based driver assistance. Waymo has partnered with Fiat Chrysler on a minivan. Waymo employs LiDAR and RADAR.
The situation is therefore risky for early movers, choosing which self-driving platform they should bet on, says Canalys’ Cox. “If we get to the point where Waymo is productized and proves to be more advanced than what other companies are capable of doing, it might be waste of investment for those companies,” he said.
What about the public? The public wants a lot...
Another factor that will decide winners and losers will be government. As officials inevitably set standards, those who fall below the bar will be forced to withdraw.
“The odds are quite long on all the horses at the moment, but who knows how it will pan out? There will be consolidation when there’s regulation, when legislators apply standards around performance of your technology,” Reed said.
Ultimately, however, there’s a bigger concern for Reed: backlash. Or, the public’s expectations are not met by what’s being promised, by the tech consortia behind autonomous vehicles and the breathless media reporting on them. There is a risk that driverless vehicles will plough head-first into a Gartner-esque trough of disillusionment before their wheels really hit the road.
“I think there’s a lot of momentum and a number of projects are making predictions and targets around certain dates. If those dates are missed and if expectations aren’t met – if the technology doesn’t deliver in some way what was promised – that will definitely lead to a backlash,” Reed said.
Integral to avoiding that backlash, therefore, is ensuring the autonomous vehicles drive goes beyond pure technology – it must work with and for people, too. It’s not a world away from the world of classic enterprise tech: designing and delivering a system users not only need but can use easily and without hurdles. Too much software failed in this, thereby earning enterprise software a hated reputation. It's expected the self-driving rollouts will be slow-moving shuttle craft that serve campus-like environments.
A detailed TRL whitepaper (here: PDF) in July expressed concern over what it called a “strong technology push for autonomous vehicles rather than a societal pull”. The report noted: “The potential benefits will only be delivered if users’ needs are catered for and society accepts and adopts automated vehicles.”
Ocado’s chief technology officer, Paul Clarke, accepts that hurdles exist, but sees the potential. “These things do get over-promised at times, but that’s why we are working with Oxbotica, a leader in their field,” Clarke said.
But can it deliver? Media hype plus a tech-only focus could damage driverless’ chances before it’s delivered. Pic: Gavin Clarke
For Ocado, success would mean another feature for OSP, for its own use – driverless warehouse vehicles and driverless deliveries – and for corporate customers.
Ocado’s goal is to sell the OSP infrastructure to other supermarkets – so far, it has just two: Morrisons Online and another, unnamed, European supermarket.
The firm built OSP when it started in 2000 because the systems of the day weren’t geared for a pure-play grocery retailer: they employed forecasting algorithms geared more to traditional brick-and-mortar businesses rather than an online firm generating and dicing lots of data.
“The fact we are trying this [self-driving] shows the depth of our technological ambitions, because we are a blend of a retailer and a technology business,” Clarke said. ®
We’ll be covering machine learning, AI and analytics – and ethics – at MCubed London in October. Full details, including early bird tickets, right here.